Instrument before you optimise
The first thing we do on an onboarding engagement is refuse to change anything. Without a step-level funnel you are redesigning by anecdote, and the loudest anecdote is rarely the biggest leak. Fire a single event per step with a stable name, the step index, and a session identifier — six to ten events is usually the whole funnel. Then add time-to-step and error events on every field that can reject input. A week of clean data on a product with a few thousand signups is enough to see the shape. On one B2B tool we did exactly this and found 41% of users never reached step three, while the team had spent two sprints polishing step five. The instrumentation cost two days. The redesign they were about to ship would have cost six weeks and moved nothing.
Find the cliff, not the slope
Funnels lose people two ways. A slope is a steady few percent at each step — annoying, structural, expensive to fix. A cliff is one step haemorrhaging 30% or more, and that is almost always a specific, cheap defect: a required field nobody has the answer to, an email verification that lands in spam, an OAuth screen requesting scopes that look alarming. Rank steps by absolute users lost, not by conversion rate, or you will optimise a 20% drop on a step only 200 people reach. Then watch ten session recordings of people who dropped at the worst step. Ten is enough; you will see the same hesitation repeat by the fourth. Cliffs are worth fixing this week. Slopes are worth a roadmap.
Progressive disclosure beats a shorter form
"Cut the form in half" is the standard advice, and it is often wrong — you still need the data, so you have moved the friction somewhere with worse visibility. Progressive disclosure works better: ask for the minimum that unlocks the next screen, and collect the rest at the moment it becomes obviously necessary. Company size belongs on the screen where it changes what the user sees, not on signup. One field per decision, and never make someone leave to find an answer. Where a field is genuinely required early, explain the cost in one line of text next to it. On the same B2B tool, splitting an eleven-field signup into three fields plus two contextual asks later moved step-three completion from 59% to 84%, with no loss in data completeness after seven days.
Defer every ask until after first value
The strongest lever is ordering. Anything you request before the user has felt the product working is a bill for a service not yet rendered — payment details, teammate invites, calendar permissions, a workspace name they have no basis to choose. Define "first value" concretely for your product: the first report rendered, the first message sent, the first repository connected. Get there in under two minutes, with seeded sample data if the real thing takes longer. Then ask. Invites converted 3.1x better for us when moved from step two to immediately after a user's first successful export, and the users who accepted were materially more likely to be active at day 30. Same request, same copy — different moment.